OpenAI announced GPT‑Live‑1 in the API on September 10, 2026, giving outside developers a full-duplex voice layer that can keep listening while speaking and handle interruptions, pauses and background noise. The release is not simply another speech-to-text, text-model and text-to-speech chain. It separates the conversational front end from backend reasoning so developers can compose the two. OpenAI lists the voice layer at $0.05 per minute; backend models and tools are charged separately.
What does GPT‑Live‑1 expose to developers?
Full-duplex means the system can receive a user’s audio and produce a response at the same time instead of forcing one side to finish before the other speaks. GPT‑Live‑1 still provides speech transcripts and response text, and supports turn detection and keyword biasing, so developers do not lose explicit turn control. The difference is that the audio interaction no longer has to feel like a push-to-talk radio.
OpenAI points to WebRTC, WebSockets and telephony as supported directions, so the target is not only a browser demo. A production service still has to handle audio streaming, call quality, transcription errors, voice consent, data retention and human handoff. The model supplies a conversation capability, not a complete contact-center or telecom-compliance package.
Conversation and action are split into layers
GPT‑Live‑1 handles the front-end voice layer for natural conversation, while deeper reasoning, data lookup and tool calls can be delegated to a backend text model such as GPT‑6 Astra or to a developer’s own model and agent. The split lets the voice model focus on listening, responding, interruptions and style, while the application owns task state, permissions and tool execution.
The benefit is that the backend can change with the task: a simple lookup does not need the most expensive reasoning every time, while complex work can use a stronger model. It also puts responsibility on the developer. When a user interrupts, the app must define whether backend work is cancelled, retried or continued. More natural audio does not automatically make tool actions safer.
$0.05 per minute covers only the voice layer
OpenAI lists the GPT‑Live‑1 front-end voice layer at $0.05 per minute, billed by the second. A ten-minute voice interaction would therefore be about $0.50 for that layer alone. It is not the full cost: the backend GPT model, function tools, telephony provider, transcription and storage may all be billed separately.
OpenAI says early Speak evaluations saw almost 80% fewer interruptions for language learners than previous turn-based systems, and it reports a 30-point improvement on its Full Duplex Bench. These are company or partner evaluations, not an independent guarantee for every use case. Before launch, teams should test their own call lengths, languages, noise conditions and tool-success rates.
Will this actually make voice products more natural?
GPT‑Live‑1’s value will depend on whether voice products can move from merely answering to completing work reliably through interruptions and waiting. For support, reservations, language learning and phone agents, less rigid turn-taking may improve the experience. For developers, the new cost model and two-layer architecture also mean more state management and safety responsibility. It is a major voice-front-end update, not automatic governance of backend tasks or human service.
